ModelMoe/UltraHR-100K-UniPercept

UltraHR-100K-UniPercept 是针对 UltraHR-100K 数据集的图片质量标注数据集。我们利用专门用于感知级图像理解的前沿多模态大模型 UniPercept，对源数据集中的 10 万张超高清图像（分辨率 > 3K）进行了深度的感知维度打标处理。本标注数据集为每张图像提供了以下三个维度的精细化评分：1. IAA (图像美学评估): 捕捉图像的艺术感与美学吸引力。2. IQA (图像质量评估): 衡量技术层面的忠实度、清晰度与失真程度。3. ISTA (图像结构与纹理评估): 评估超高清场景下细粒度细节的丰富度与结构完整性。

UltraHR-100K-UniPercept is an image quality annotation dataset built upon the UltraHR-100K dataset. We utilized the cutting-edge multimodal large model UniPercept, which is specifically designed for perceptual-level image understanding, to perform in-depth perceptual dimension-based annotation on 100,000 ultra-high-resolution images with resolution exceeding 3K from the source dataset. This annotated dataset provides refined scoring across three dimensions for each image:
1. IAA (Image Aesthetic Assessment): Captures the artistic quality and aesthetic attractiveness of the image.
2. IQA (Image Quality Assessment): Quantifies technical fidelity, clarity, and the degree of image distortion.
3. ISTA (Image Structure and Texture Assessment): Evaluates the richness of fine-grained details and structural integrity in ultra-high-resolution scenarios.